What happened
OpenCode announced that DeepSeek V4.1 Flash is now available free on its platform. The report says that the free version supports a 200K context window, which is lower than the 1M-context version referred to in the report. More importantly for an account holder, calls through the newly free offering no longer count against usage quota. The article does not give an end date for the free access, nor does it set out a separate rate limit, eligibility rule, or geographic availability condition.
The move is presented as a further development of OpenCode’s earlier “Operation Cheepseek” plan. In September, OpenCode had temporarily raised the DeepSeek V4.1 Flash allowance in its OpenCode Go plan from $15 to $60. The report says that on October 12 the allowance was then adjusted to $80. The new offer changes the mechanism again: rather than providing a larger dollar allowance inside the plan, calls to the model on OpenCode do not consume that allowance. The report does not say whether every capability available to paid users is included in the free path.
DeepSeek V4.1 Flash itself was formally released on September 10, 2026, the article says. It describes the model as a natively multimodal mixture-of-experts model with 552B backbone parameters and 196B Engram conditional-memory parameters. Its architecture is described as a causal encoder-decoder: the first 20 layers of a 40-layer language backbone are a causal encoder, while the remaining 20 are a decoder. The report says this design brings prefill compute close to a halving. It also says that DeepSeek’s own platform has no free API tier and uses prepaid, usage-based billing.
The 200K cap is only one of the operational constraints surfaced in the report. Community discussion has focused on generation throughput, measured as tokens per second. One user reported observing output as high as roughly 490 tokens per second on the Ollama cloud platform, but also described substantial instability, with speed dropping to roughly 50–60 tokens per second from midday into the afternoon. Within OpenCode Go, users reported throughput falling below 100 tokens per second. Other community tests reported that speed declines as context grows: below 100 tokens per second after 100K context, and around 50 tokens per second at 200K context. These are user reports, not a stated service-level commitment or controlled benchmark.
The immediate result is therefore specific. A builder can invoke DeepSeek V4.1 Flash in OpenCode without spending plan quota, but cannot assume access to the model’s full reported 1M context window. Nor can that builder infer stable throughput from the free-access announcement. The public report is available at https://www.ithome.com/1/013/475.htm.
What the archive adds
This is a new access change on a known OpenCode and DeepSeek V4.1 Flash story. Our earlier note on DeepSeek V4.1 Flash usage, cost and rank argued that downstream platform usage is an adoption signal, not a proxy for quality or broad market share. A separate note recorded OpenCode’s permanent $60 quota arrangement under Operation Cheepseek Phase 2: OpenCode makes $60 DeepSeek V4.1 Flash quota permanent. The reported free, non-quota calls are a different access condition.
Trend signal
No supplied pattern directly fits: a quota-free model offer changes access and cost, but the material does not establish a shift in completed-work economics, protocol adoption, or runtime architecture.
Why it matters
Free access lowers the cost of trying a model inside an existing coding-agent workflow. That matters most when evaluation has been blocked by quota anxiety: a team can run representative tasks, inspect outputs, and compare model behavior without each experiment depleting its OpenCode allowance. But the offer changes neither the need for an evaluation harness nor the physical constraints of a long-running task.
A 200K limit can be ample for bounded changes, focused repository work, or isolated tool loops. It is not equivalent to a million-token workspace. The reported throughput behavior creates a second planning variable: as context expands, an agent may become slower even before it reaches the cap. Teams should treat the offer as an inexpensive test surface, not as evidence that long-context production work will be predictably fast or unrestricted.
Limits and uncertainty
The account is a report of an OpenCode announcement and community testing, rather than a published service specification in the supplied material. It does not establish the duration of free access, eligibility, rate limits, concurrency limits, supported OpenCode features, or whether the 200K cap applies identically to every request type. The speed figures are anecdotal observations across platforms and times, not independently controlled measurements. The model’s parameter and architecture descriptions are reported facts, but the material does not demonstrate comparative task quality, reliability, safety, or suitability for a particular codebase.
Skill to build now
Create a small evaluation pack for one real coding workflow this week: include a short-context task, a task that approaches your usual repository-context size, and a tool-using task with a verifiable outcome. Record context size, elapsed time, output tokens, successful completion, and any manual repair. Run it through the free OpenCode route while it is available. This separates the useful fact—whether the model completes your work—from a headline about free calls or a single tokens-per-second anecdote.
Reader takeaway
OpenCode’s reported change makes DeepSeek V4.1 Flash available without consuming quota, which is meaningful for hands-on evaluation. The trade-off is concrete: the free route is reported to stop at 200K context, far below 1M, and community reports indicate speed can fall sharply at large contexts. Test it on bounded work before relying on it for long agent runs.
Open question
Will OpenCode publish durable terms for the free route, including its end date, rate limits, eligibility, and whether the 200K context cap is fixed?