RegulationRules, enforcement & market structure
Minerva co-founder outlines why U.S. AI startups fine-tune Chinese open-weight models, citing Harvey's Kimi K3-based Tenet
On Bloomberg's Odd Lots, Minerva co-founder Daniel Saedi describes a split: closed frontier models for building products, fine-tuned Chinese open-weight models like Qwen and Kimi K3 for repeated tasks. Harvey's Tenet legal model, post-trained on Kimi K3, is the leading example.
From the Odd Lots podcast episode of November 20, 2026 and cited sources: Saedi says his company uses closed frontier models while building its product and fine-tuned open-weight models for repeated production tasks. [1] [2]
SOURCES[1] How an American AI Company Uses Chinese Open-Weight Models omny.fm[2] Transcript: How an American AI Company Uses Chinese… api.omny.fm
AI-assisted, not human-edited. Methodology Corrections Who owns this site
- What happened
- Saedi says his company uses closed frontier models while building its product and fine-tuned open-weight models for repeated production tasks. [1] [2]
- Original source
- Odd Lots podcast episode, independent perspective
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- November 20, 2026 podcast episode published
- Covered
- Nov 20 edition
- Evidence
- 14 sources across 12 sites; transcript sourced
- 1How an American AI Company Uses Chinese Open-Weight Modelsomny.fm
- 2Transcript: How an American AI Company Uses Chinese…api.omny.fm
- 3Harvey: Post-training update, Harvey Tenetharvey.ai
- 4MarTech Series: Minerva launchmartechseries.com
- 5The Next Web: Harvey's first in-house modelthenextweb.com
- 6PlayCISO: DeepSeek V4.1 Flash Abliteratedplayciso.com
Reach: not measured yet

Saedi says his company uses closed frontier models while building its product and fine-tuned open-weight models for repeated production tasks. He ranks the reasons as cost first, measured per completed task rather than per token, then control over data, then leverage in pricing talks with frontier vendors. As an example, Minerva fine-tuned Qwen on a few thousand graded pairs of plain-language requests and correct SQL. He says the result was faster than closed models for audience building. [1] [2]
Harvey post-trained Kimi K3 with Fireworks to create Tenet. The setup used a rank-64 LoRA adapter, GSPO reinforcement learning across about 1,750 legal-agent environments, and roughly 150 B300 GPUs for two months. On Harvey's own Legal Agent Benchmark, Tenet passed 19.7% of tasks against 10.8% for base Kimi K3, at about the same cost per task ($5.92 vs $5.62). No one has independently reproduced these results. [3] [11] [12] [8]
Saedi calls 'abliterated' open-weight models, which have their safety refusals removed, the main risk of releasing open weights. Reportedly, such copies of DeepSeek V4.1 Flash appeared within hours of its launch. A ThreatDown count found 6,644 Hugging Face models labeled this way, with publishing roughly tripling from Q4 2025 to Q2 2026. Saedi's link between these models and current hacks is an internet theory. No confirmed incident has been cited. [1] [2] [6] [13]
Harvey's results suggest post-training an open-weight base can nearly double task success without raising per-task cost.
Read the full assessment
Implication: firms with fixed seat pricing and rising usage may adopt in-house models to cut costs or to negotiate with frontier vendors.
Writing up the dossier now that the main claims are checked.
Executive brief
Harvey is a legal-AI company backed by OpenAI. Its first in-house legal model, Tenet, is not built on OpenAI's models. It is post-trained on Kimi K3, an open-weight model from Moonshot AI in Beijing (Harvey). On Bloomberg's Odd Lots, Minerva co-founder Daniel Saedi uses that example to describe a common split. Teams use closed frontier models to build products, then fine-tune Chinese open-weight models to run repeated tasks cheaply. He gives three reasons in order: cost, control over data, and bargaining power with vendors (Odd Lots, 12:14–14:36). He also names the main risk: "abliterated" open models with their safety refusals removed.
What changed and event timeline
Minerva launches
The New York marketing-agent startup, co-founded by ex-Bridgewater trader Daniel Saedi, comes out of stealth with a $20M seed round and a collaboration with OpenAI ().
Harvey previews Tenet
Harvey post-trained Kimi K3 with Fireworks using asynchronous reinforcement learning (RL). It says no customer data was used (;).
DeepSeek V4.1 Flash ships
It was released under an MIT license. Reportedly, "abliterated" copies with refusals stripped out appeared on Hugging Face within hours ().
Reflection announces Beam
Reflection is backed by Nvidia. Beam is an open-weight mixture-of-experts (MoE) model with 501B total parameters, 23B of them active per token, released under Apache 2.0 ().
Odd Lots episode published
The hosts say it was recorded on Nov 3. Saedi argues that open weights plus a company's own data can produce a defensible margin ().
Capabilities and access
- Kimi K3: a 2.8-trillion-parameter model with downloadable weights (Kingy AI). LoRA training costs $32.55 per million training tokens (Fireworks pricing).
- DeepSeek V4.1 Flash: $0.30 input / $1.20 output per million tokens on Fireworks (Fireworks pricing).
- No weights or standalone API have been released (Kingy AI).
Read the full section
- Kimi K3: a 2.8-trillion-parameter model with downloadable weights (Kingy AI). Serverless pricing on Fireworks is $3 per million input tokens and $15 per million output tokens. LoRA training costs $32.55 per million training tokens (Fireworks pricing).
- DeepSeek V4.1 Flash: $0.30 input / $1.20 output per million tokens on Fireworks (Fireworks pricing).
- Harvey Tenet: a research preview only. No weights or standalone API have been released (Kingy AI).
- Reflection Beam: a 1M-token context window, with weights, a technical report and fine-tuning tools promised for late October (DeepLearning.AI).
Technical analysis for researchers and developers
- How Tenet was trained:
- A rank-64 LoRA adapter covering the attention, MLP and routed-expert weights. LoRA trains a small add-on layer instead of the full model.
- Group-sequence policy optimization (GSPO), an RL method, run across about 1,750 legal-agent environments.
Read the full section
- How Tenet was trained:
- A rank-64 LoRA adapter covering the attention, MLP and routed-expert weights. LoRA trains a small add-on layer instead of the full model.
- Group-sequence policy optimization (GSPO), an RL method, run across about 1,750 legal-agent environments.
- About 150 B300 GPUs for two months (ZenML).
- Rewards favored solutions that used fewer tokens when results were equal (Harvey).
- Minerva's approach: a few thousand graded pairs of plain-language requests and correct SQL, used to fine-tune Qwen for building audiences (Odd Lots, 09:08).
- Durability: a fine-tune does not carry over when the base model is upgraded. The dataset can be reused, but whether retraining reaches the same quality is uncertain (Odd Lots, 14:54).
Claims and evidence
- Harvey's results:
- Tenet passes 19.7% of tasks on Harvey's Legal Agent Benchmark (LAB), against 10.8% for base Kimi K3.
- Cost per task is about flat: $5.92 versus $5.62 (NxCode).
Read the full section
- Harvey's results:
- Tenet passes 19.7% of tasks on Harvey's Legal Agent Benchmark (LAB), against 10.8% for base Kimi K3.
- Cost per task is about flat: $5.92 versus $5.62 (NxCode).
- Harvey wrote and ran LAB itself, and no one has independently reproduced the results (Kingy AI).
- Harvey's spend: Saedi's figure of about $6M in compute does not appear in Harvey's post, which gives no total spend (Harvey).
- Fine-tuning prices: Saedi cites $7–9 per million tokens. Current Fireworks list prices for Kimi K3 are higher. Rates vary by model size (Fireworks pricing).
- Reflection: Saedi says Beam trained on 12,000–14,000 GPUs. Reporting says 6,144 GB300s for pretraining and 10,500 for RL (The Next Web). Nvidia is reported to be only "in early talks" about a deeper deal (DeepLearning.AI).
- Cognition: no corroboration was found for Saedi's statement that Cognition fine-tunes Kimi.
Context and prior work
- Saedi says open-source releases include the code, data and training recipe, while Chinese labs release only the weights (Odd Lots, 07:31).
- U.S. open-weight models still trail Chinese ones on one widely used index:
- Beam scores 30 on the Artificial Analysis index.
Read the full section
- Saedi says open-source releases include the code, data and training recipe, while Chinese labs release only the weights (Odd Lots, 07:31).
- U.S. open-weight models still trail Chinese ones on one widely used index:
- Beam scores 30 on the Artificial Analysis index.
- The top U.S. open model, Meta's Muse Spark Open, scores 39.
- The top Chinese models, Xiaomi MiMo-V2.6-Pro and Z.ai GLM-5.3, score 46 and 45 (DeepLearning.AI).
- On Reflection's own Terminal Bench numbers, Beam scores 80.1 against Kimi K3's 88.3 (The Next Web).
- Saedi uses that gap to argue that distillation (training on outputs from Western models) helps Chinese labs but does not explain all of their progress (Odd Lots, 28:31).
Limitations, safety and contested findings
- Abliterated models: ThreatDown counted 6,644 Hugging Face models labeled "abliterated," "uncensored" or "heretic."
- Hacking claims: Saedi says swarms of abliterated models are behind current hacks, but frames it as an internet theory (Odd Lots, 24:45).
- How much of Chinese models' quality it explains is still disputed.
Read the full section
- Abliterated models: ThreatDown counted 6,644 Hugging Face models labeled "abliterated," "uncensored" or "heretic." Publishing of such models roughly tripled between Q4 2025 and Q2 2026 (Morph).
- Hacking claims: Saedi says swarms of abliterated models are behind current hacks, but frames it as an internet theory (Odd Lots, 24:45). Morph's write-up cites no confirmed incident.
- Distillation: Saedi calls it real. How much of Chinese models' quality it explains is still disputed.
- Licensing: Kimi K3's terms require a separate agreement with Moonshot for operators earning more than $20M a year. Harvey's annualized revenue is above $350M (The Next Web).
Business and practitioner implications
- Unit economics: Saedi argues that the useful measure is cost per completed task, not price per token.
- Bargaining power: an in-house model gives leverage in pricing talks with Anthropic and OpenAI, even if it is never served to customers.
- Infrastructure: hosted fine-tuning has removed the need for a dedicated team of inference engineers. Fireworks reports $1B in annual recurring revenue (Fireworks).
Read the full section
- Unit economics: Saedi argues that the useful measure is cost per completed task, not price per token. He suggests swapping in open models where fixed seat-based revenue meets rising usage costs (Odd Lots, 34:22).
- Bargaining power: an in-house model gives leverage in pricing talks with Anthropic and OpenAI, even if it is never served to customers.
- Infrastructure: hosted fine-tuning has removed the need for a dedicated team of inference engineers. Fireworks reports $1B in annual recurring revenue (Fireworks).
- Diligence: check license revenue thresholds, the cost of retraining with each new base model, and compliance under the EU AI Act.
The source trail.
Sources (14)
How an American AI Company Uses Chinese Open-Weight Models
Transcript retrieved via published_transcript; language en. Timestamped text, not direct audiovisual review. Source: https://api.omny.fm/orgs/e73c998e-6e60-432f-8610-ae210140c5b1/clips/4364f8cd-e957-44ef-9795-b5070115b19e/transcript?format=TextWithTimestamps&t=1795107325. Automatic captions/transcription may contain errors.
omny.fmTranscript: How an American AI Company Uses Chinese Open-Weight Models
published_transcript
api.omny.fm